V o L O ( V o x e l - L e s s O p t i m i z at i o n )

T e c h n olo g y

(Latin: Volo = I fly)

V o L O ( V o x e l - L e s s O p t i m i z a t i o n ) T e c h n olo g y

The new Accuray VoLOTM technology represents an exciting step forward for the TomoTherapy treatment planning
process, and promises to make treatment planning faster, more flexible and more interactive, making optimal
TomoTherapy treatments available to more patients. It combines the power of graphics processing units (GPUs)
traditionally used in the computer gaming world with a new Non-Voxel Broad Beam (NVBB) algorithm, an innovative
plan optimization framework.

The NVBB framework is the result of re-engineering the TomoTherapy Systems plan optimization algorithm, altering
the way in which beam and patient geometries are visualized and also eliminating the need for creation and storage of
large amounts of data. This has made it possible to take advantage of the raw processing power of the GPU platform
whose strength is non-memory intensive yet highly-parallelized computations.

VoLO Technical Attributes

The key properties of VoLO can be summarized as follows:

GPU implementation using hundreds of processor cores for highly-parallelized, ray-by-ray dose computations
and beam element updates
Continuous, non-voxel broad beam (NVBB) representation of beam and patient geometry
Collapsed cone convolution superposition (CCCS) combined with fluence convolution broad beam (FCBB)
dose calculations for accuracy and speed
Use of FCBB during iterations eliminates the need for a beamlet calculation preprocessing step
Direct treatment parameter optimization (DTPO) for determination of machine control parameters
A high degree of flexibility in altering machine parameters during optimization

GPU Implementation and the NVBB Framework

Each GPU card used by the VoLO system contains 448 processors (cores) that work simultaneously to calculate dose
and optimize the parameters that control the treatment delivery system. There are two cards per GPU node mounted
in the data server/optimizer assembly. The standard configuration (single node) enables planning on up to four
TomoTherapy Planning Stations at one time. The high performance configuration (two nodes) enables planning on
up to five Planning Stations at one time. This simultaneous usage includes both dose calculation and optimization.

Figure 1. Schematic of a Fermi GPU card. Each small square represents

a single processing unit.
Parallelization is suited to computational problems that involve a very large number of discrete elements, whose
properties need updating rapidly to contribute to the overall result. While GPU implementation of a radiotherapy
optimization problem is ideal in principle, data storage requirements must be minimized due to the limited
memory available in the GPU architecture. The NVBB computational geometry and optimization approach used
in the VoLO technology were specially adapted for a GPU-based implementation.

The NVBB framework is very economical in memory utilization because it does not involve storage of pre-computed
beamlet dose distributions to be used during iterative optimization. Instead, innovative on-the-fly dose calculations
are used for the large number of 3D dose evaluations that are necessary during optimization. These calculations are
accurate, fast and have minimal memory overhead.

In the NVBB framework, a beams eye view perspective of the patient geometry is used. Dose calculations and
optimization of system parameters are done by considering rays diverging from the trajectory of beam source
locations during either a TomoHelical or TomoDirect delivery. One such diverging beam geometry is shown
in Figure 2. Rays intersect a plane perpendicular to rays joining the source and machine isocenter, known as the
beams eye view plane, on which beam fluence entering the patient is defined. The interaction of these rays with
the patient geometry and elements of the delivery system such as the jaws and the leaves of the binary multi-leaf
collimator are inherent in the algorithm. A direct treatment parameter optimization (DTPO) approach is used to
update machine parameters. For further reading see reference [1].

The specific dose calculation algorithms used

within the NVBB framework are known as fluence
convolution broad beam (FCBB) and collapsed
cone convolution superposition (CCCS). These Continuous broad beam geometry
are used for different aspects of the optimization
process.

As shown in Figure 3, the NVBB framework

involves full dose, iteration dose and final dose
calculations. In addition, an adaptive full dose
correction step is used as part of each iteration
Beams eye view plane
dose step to increase accuracy.

The fluence convolution broad beam algorithm

is used during the iteration dose steps. This
algorithm is designed to allow rapid 3D dose Dose computation
calculations, while making certain time-saving along diverging rays
approximations about radiation transport in
heterogeneous material. For further reading
see reference [2].

The collapsed cone convolution superposition

algorithm is used during the full dose and final
dose steps. This 3D dose calculation algorithm Figure 2. Diverging rays from a source position intersecting a beams
is more rigorous than FCBB in modeling radiation eye view plane. Dose is calculated at points along each ray.
transport in heterogeneous material.

The adaptive full dose correction employs known differences between FCBB and CCCS to make an additive
correction to each iteration dose. The difference is determined every tenth iteration and improves the
accuracy of other iterations without the need for a full dose calculation. This is a key element of the overall
optimization strategy.

When the user is satisfied with the plan result, a final dose is calculated that fully takes into account constraints
of the delivery system and sets delivery parameters according to the number of fractions chosen for the treatment
course. Machine parameters used for system control are then stored ready for patient quality assurance and
subsequent delivery.

2
V o L O ( V o x e l - L e s s O p t i m i z a t i o n ) T e c h n olo g y

Non-Voxel Broad Beam (NVBB) Framework

Iteration doses with

Full adaptive full dose Final
dose correction dose

Initial machine Final machine

parameters parameters

Optimize until clinical

Direct treatment
objectives met
parameter optimizations

Figure 3. Process used for optimization in the NVBB framework. After initial machine parameters are used as inputs, direct treatment parameter
optimization is performed iteratively until a satisfactory dose distribution is attained. Full dose calculated every tenth iteration is used in an
adaptive full dose correction during other iterations.

Clinical and Operational Advantages

Interactivity and flexibility are the two key attributes of the VoLO planning process. In particular, the VoLO
technology provides:

Fast overall planning time, especially for cases involving large or complex treatment volumes. TomoTherapy
planning for large and complex geometries was always conceptually simple, but now high speed is added to
simplicity for all cases.
Optimization that can begin as soon as contouring is completed and essential machine parameters are defined
by the user.
Extra flexibility to explore alternative parameters such as field width, helical pitch and dose grid resolution
interactively. Because machine parameter-specific beamlets are not required with NVBB, options for improving
plan quality can be investigated without incurring time overheads.

These advantages bring the potential to plan more patients faster, thereby making optimal TomoTherapy
treatments available to additional patients within the clinic.

Efficiency

In the peer-reviewed paper describing the NVBB algorithm [1] a simple planning efficiency test was performed for
a range of Helical TomoTherapy cases with different levels of complexity, number of dose calculation points and
number of beamlets. Here, number of beamlets refers to the number of open MLC (multileaf collimator) leaf
instances occurring during the treatment delivery. This number ranges from approximately 4,700 for the prostate
case to 117,000 for the total marrow irradiation (TMI) case. In previous versions of the TomoTherapy optimizer,
a 3D dose distribution would have been created and stored for each beamlet during the preprocessing step,
typically requiring from several minutes to several tens of minutes. Preprocessing is virtually eliminated with the
NVBB implementation. For these examples, the number of dose calculation points ranges from approximately
2 million to 8 million, depending on the irradiated volume and dose grid resolution.

Cumulative time components are shown in Figure 4 for each case. Durations are shown for (i) preprocessing, (ii)
100 optimization iterations, (iii) full dose, and (iv) final dose. Note that iteration times include full dose calculations
as part of the adaptive full dose correction process. These plots are simple indications of planning time and would
be affected by human interaction in a clinical setting. As an indication of how these times compare with the previous
CPU-based optimization strategy including beamlet calculations, the breast case and TMI cases had planning times
of approximately 30 minutes and 150 minutes respectively. This represents a time reduction of 10-20 times, mostly
due to the elimination of the beamlet calculation step. It is clear that preprocessing, full dose and final dose are not
significant time components with this new strategy. Iteration times are fast also due to the extreme parallelization
of the computation process via GPU implementation. Note that the relative variability in overall planning time is
surprisingly small considering the large variation in number of beamlets and dose calculation points. The variation
in absolute time in minutes is also small. Cases traditionally considered complex can be planned quickly using
this new framework.
 Planning Time Components
3.2m dose points
117,000 beamlets
TMI

8.2m dose points

H&N
8,100 beamlets

3.2m dose points

Breast 14,600 beamlets

2.0m dose points Preprocessing

Lung 11,800 beamlets 100 Interations
Full dose
3.1m dose points Final dose
Prostate 4,800 beamlets

0 2 4 6 8 10

Minutes

Figure 4. Planning time components for a range of cases including preprocessing, 100 optimization iterations, full dose calculation and final
dose calculation Note that the 100 iterations step includes a full dose calculation every ten iterations as part of the adaptive full dose correc-
tion process. Number of beamlets (MLC leaf opening events) and dose calculation points are indicated for each case. Data from reference [1].

Extra insight into the performance improvements brought about by the VoLO technology can be gained via a com-
parison of optimization times with those for previous generations of the TomoTherapy optimizer. Shown in Figure 5
are the times for a 200-iteration optimization session for a head and neck cancer case for VoLO along with four older
generation systems. Gen 5 is the Generation 5 CPU-based hardware currently shipped as the standard configuration
on all HiArt Systems. The VoLO (labeled GPU) technology can be seen to be approximately three times as fast as
Gen 5 for this case and more than ten times faster than optimization using original TomoTherapy hardware. Except for
the GPU case, times include the pre-optimization beamlet calculation step that is unnecessary with the VoLO technology.

Some of the time reduction seen with the VoLO technology is due to the absence of the beamlet calculation step,
some is due to the parallelization inherent in GPU operation, and some is due to other increases in algorithm efficiency.

Optimization Times for Different Hardware Generations

(Head & Neck Case 200 iterations)

GPU

Gen 5

DC3

DC2

DC1

0 20 40 60 80 100 120 140

Minutes

Figure 5. Optimization times for VoLO (GPU) versus previous generations of the TomoTherapy optimizer for an example head and neck
cancer case. Gen 5 is the Generation 5 CPU-based hardware currently shipped as the standard configuration on all HiArt systems.
Each time plotted includes 200 optimization iterations. Except for GPU, times include a pre-optimization beamlet calculation step.

3
V o L O ( V o x e l - L e s s O p t i m i z a t i o n ) T e c h n olo g y

Further comparing VoLO planning speed with that of the generation 5 hardware, Figure 6 shows optimization
times for breast, head and neck and craniospinal cases. Again, 200 iterations were performed for each case.
The head and neck case is the same one used to generate the data in Figure 5. The breast example used the
TomoDirect delivery technique so the total number of beamlets employed is much less than if this had been
a TomoHelical delivery. As with the head and neck case, optimization times for the breast and craniospinal plans
are reduced by approximately a factor of three between Gen 5 and GPU.

Optimization Times for Different Cases (200 iterations)

(GPU and Generation 5 Hardware)

Breast (TomoDirect)

Head & Neck

GPU
Gen 5

Craniospinal

0 10 20 30 40 50 60

Minutes

Figure 6. Optimization times for breast, head and neck and craniospinal cases for the VoLO (GPU) and Generation 5 hardware. 200 iterations
were performed for each case. The head and neck case is the same one used to generate the data in Figure 5. For Gen 5, these times include
the pre-optimization beamlet calculation step that is unnecessary with the VoLO technology.

Although a large contribution to the increase in speed between Gen 5 and VoLO is the absence of an initial beam-
let calculation, optimization iterations and other dose calculations are also faster. Figure 7 shows a comparison of
average individual iteration time for the same three cases. In the case of VoLO, iteration time is an average over ten
iterations where one is a full dose calculation. As described above, one in every ten iterations is a full Collapsed
Cone Convolution Superposition (CCCS) calculation and the others use the more approximate Fluence Convolution
Broad Beam (FCBB) algorithm.

Average Individual Iteration Times for Different Cases

(GPU and Generation 5 Hardware)

Breast (TomoDirect)

Head & Neck

GPU
Gen 5

Craniospinal

0 1 2 3 4 5 6

Seconds

Figure 7. Average individual iteration times for the VoLO (GPU) and Gen 5 configurations for the same cases as in Figure 6. This illustrates that
reduction in iteration time contributes to the overall increase in optimization speed.
Finally, Figure 8 shows the reduction in final dose computation times between Gen 5 and GPU configurations.
For the head and neck plan as well as the craniospinal plan this graph shows a more dramatic difference than for
the iteration times. Final dose is computed once at the end of the plan optimization process. This calculation step
takes account of practical machine characteristics and creates final instructions for machine control per fraction
of the treatment course. Time includes saving of the dose and delivery instructions to the central data server. The
dramatic reduction in time is mostly due to the extreme parallelization of the algorithm enabled by the use of a GPU.

Final Dose Calculation Times for Different Cases

(GPU and Generation 5 Hardware)

Breast (TomoDirect)

Head & Neck

GPU
Gen 5

Craniospinal

0 50 100 150 200 250

Seconds

Figure 8. Average final dose calculation times for the VoLO (GPU) and Gen 5 configurations for the cases in Figure 6 and Figure 7. This cal-
culation step takes account of practical machine characteristics and creates final instructions for machine control per fraction of the treatment
course. Time includes saving of the dose and delivery instructions to the central data server.

To close the loop with this new treatment planning platform, rigorous tests of its accuracy have been carried out,
including consistency with the traditional CPU-based platform used by current clinical systems. Agreement with
measured doses, highly-accurate Monte Carlo dose calculations, as well as the CPU platform has been shown to
be excellent [3].

References

1. Lu W. A non-voxel-based broad-beam (NVBB) framework for IMRT treatment planning. Phys Med Biol. 2010
Dec 7;55(23):7175-210.
2. Lu W, Chen M. Fluence-convolution broad-beam (FCBB) dose calculation. Phys Med Biol. 2010 Dec
7;55(23):7211-29.
3. Chen Q, Lu W, Chen Y, Chen M, Henderson D, Sterpin E. Validation of GPU based TomoTherapy dose
calculation engine. Med Phys. 2012 Apr;39(4):1877-86.

4
UNITED STATES ASIA EUROPE

Accuray Corporate Headquarters TomoTherapy Accuray Japan K.K. - Tokyo Accuray Asia Ltd. - Hong Kong Accuray Chengdu Accuray Europe Tomotherapy Belgium BVBA
1310 Chesapeake Terrace 1240 Deming Way Shin Otemachi Building 7F Suites 1702 - 1704, Tower 6 #39 Huatai Road, Longtan Tour Atlantique 25e Pegasuslaan 5
Sunnyvale, CA 94089 Madison, WI 53717 2-2-1 Otemachi, Chiyoda-ku The Gateway, Harbour City Industrial Zone, 1 Place de la Pyramide 1831 Diegem
USA USA Tokyo 100-0004 9 Canton Road, T.S.T. Section 2 East, 3rd Ring Road, 92911 Paris La Dfense Cedex Belgium
Tel: +1.408.716.4600 Tel: +1 608 824 2800 Japan Hong Kong Chengdu, Sichuan 610051 France Tel: +32 (0)2 400 4400
Toll Free: 1.888.522.3740 Fax: +1 608 824 2996 Tel: +81.3.6265.1526 Tel: +852.2247.8688 China Tel: +33.1.5523.2021 Fax: +32 (0)2 400 4401/02
Fax: +1.408.716.4601 Fax: +81.3.3272.6166 Fax: +852.2175.5799 Fax: +33.1.5523.2039
Email: sales@accuray.com

2012 Accuray Incorporated. All Rights Reserved. Accuray, the stylized logo, CyberKnife, Synchrony, Xsight, Xchange, RoboCouch, TomoTherapy, HiArt, and TomoHD are among the trademarks and/or registered trademarks of Accuray Incorporated
in the United States and other countries. TomoTherapy is a wholly owned subsidiary of Accuray. 500917.A

UNITED STATES ASIA EUROPE

Accuray Corporate Headquarters TomoTherapy Accuray Japan K.K. - Tokyo Accuray Asia Ltd. - Hong Kong Accuray Chengdu Accuray Europe Tomotherapy Belgium BVBA
1310 Chesapeake Terrace 1240 Deming Way Shin Otemachi Building 7F Suites 1702 - 1704, Tower 6 #39 Huatai Road, Longtan Tour Atlantique 25e Pegasuslaan 5
Sunnyvale, CA 94089 Madison, WI 53717 2-2-1 Otemachi, Chiyoda-ku The Gateway, Harbour City Industrial Zone, 1 Place de la Pyramide 1831 Diegem
USA USA Tokyo 100-0004 9 Canton Road, T.S.T. Section 2 East, 3rd Ring Road, 92911 Paris La Dfense Cedex Belgium
Tel: +1.408.716.4600 Tel: +1 608 824 2800 Japan Hong Kong Chengdu, Sichuan 610051 France Tel: +32 (0)2 400 4400
Toll Free: 1.888.522.3740 Fax: +1 608 824 2996 Tel: +81.3.6265.1526 Tel: +852.2247.8688 China Tel: +33.1.5523.2021 Fax: +32 (0)2 400 4401/02
Fax: +1.408.716.4601 Fax: +81.3.3272.6166 Fax: +852.2175.5799 Fax: +33.1.5523.2039
Email: sales@accuray.com

2012 Accuray Incorporated. All Rights Reserved. Accuray, the stylized logo, CyberKnife, Synchrony, Xsight, Xchange, RoboCouch, TomoTherapy, HiArt, and TomoHD are among the trademarks and/or registered trademarks of Accuray Incorporated
in the United States and other countries. TomoTherapy is a wholly owned subsidiary of Accuray. 500917.A